PEDL

PEDL is a method for predicting protein-protein assocations from text. The paper describing it will be presented at ISMB 2020.

Requirements

• python >= 3.6
• pip install -r requirements.txt
• pytorch >= 1.3.1 (has to be installed manually, due to different CUDA versions)

Generate data

We use two types of data sets: Data generated from the BioNLP-ST event extraction data sets and the distantly supervised PID data set

Generate BioNLP

./conversion/make_bionlp_data.sh generates the BioNLP data sets for both PEDL and comb-dist

All experiments in the paper have been performed with the masked version of the data, e.g. distant_supervision/data/BioNLP-ST_2011/train_masked.json.

Generate PID

Generating the PID data is a bit more involved:

1. First, we have to download the raw PubMed Central texts: python download_pmc.py. CAUTION: This produces over 200 GB of files and spawns multiple processes.
2. Then, we have to download the PubTator Central file (ftp://ftp.ncbi.nlm.nih.gov/pub/lu/PubTatorCentral/bioconcepts2pubtatorcentral.offset.gz) and place it into the root directory. This file consumes another 80 GB when decompressed.
3. Generate the raw PID data: ./conversion/generate_raw_pid.sh
4. Generate the final PID data: ./conversion_make_pid.sh

Training PEDL

Before training, SciBERT has to be downloaded and placed to some directory (called $bert_dir from now on).

The vocabulary of SciBERT has to be adapted to include the entity markers and protein masks: cp distant_supervision/vocab.txt $bert_dir

PEDL can be trained with python -m distant_supervision.train_pedl, (see train_pedl.sh for exact suitable arguments.

If you just want to reproduce the experiments from the paper, this can be achieved with ./train_pedl.sh.

Pretrained model

As an alternative to training your own model, you can use this version of PEDL that was trained on PID and used for the experiments in the paper.

Predicting with PEDL

The trained PEDL model can be used to predict PPAs for a new data set. See predict_pedl.sh for details.

Disclaimer

Note, that this is highly experimental research code which is not suitable for production usage. We do not provide warranty of any kind. Use at your own risk.